@agentix-e/nl2spel

Natural Language → SpEL Core Engine

Zero external deps · Four-layer hybrid architecture · 532 tests · Chinese & English

API Documentation

Full API documentation → nl2spel API

Core Modules

import {
  NL2SpelEngine,        // Main engine
  PatternMatcher,       // Layer 0: Pattern matching (63 built-in patterns)
  IntentClassifier,     // Layer 1: Intent classification (15 NLIntent types)
  TemplateEngine,       // Layer 1: Template filling (15 template categories)
  PromptBuilder,        // Layer 2: LLM Prompt construction
  ValidationPipeline,   // Validation pipeline (4 stages)
  AutoFixer,            // Auto-fixer (10 error types)
  SelfCorrectionLoop,   // Self-correction loop (max 3 iterations)
  StrategyRouter,       // Strategy router (Pattern → Template → LLM)
  ContextExtractor,     // Context extraction
  SchemaFormatter,      // Schema formatting
  ChineseNumberParser,  // Chinese number parsing
  ProviderRegistry,     // Provider registration management
} from '@agentix-e/nl2spel';

Architecture

NL2SpelEngine
├── StrategyRouter          (Three-layer strategy routing)
│   ├── Layer 0: PatternMatcher    (63 built-in patterns, P99 < 1ms)
│   ├── Layer 1: TemplateEngine    (15 intents, 15 template categories)
│   └── Layer 2: LLMProvider       (Pluggable, via ProviderRegistry)
├── ValidationPipeline      (Parse → Type → Semantic → Context)
├── AutoFixer               (=== → ==, && → and, paren fixes...)
├── SelfCorrectionLoop       (Max 3 correction iterations)
├── ContextExtractor         (Auto-extract ContextSchema)
└── PromptBuilder            (Chinese/English System Prompt + EBNF + Few-Shot)

Quick Start

import { NL2SpelEngine } from '@agentix-e/nl2spel';

const engine = new NL2SpelEngine();

// Offline mode (no LLM dependency)
const result = await engine.generate('Order amount greater than 1000', { offlineOnly: true });
console.log(result.expression);  // #order.amount > 1000
console.log(result.strategy);    // pattern

Performance

| Metric | Value | |------|------| | Pattern match P99 | < 0.5ms | | Template generation | < 5ms | | Intent classification | < 2ms | | Validation pipeline | < 5ms |

License

Built on @agentix-e/spel-ts — pure TypeScript SpEL evaluator.

Ecosystem

• @agentix-e/spel-ts — SpEL parser and evaluator
• @agentix-e/spel-editor — Web-embeddable SpEL editor

MIT © 2025 Agentix-E